As generative AI adoption reshapes the competitive

landscape, businesses without a strong data strategy

will find their ambitions limited.

Produced in partnership with

Data strategies
for AI leaders
2  MIT Technology Review Insights

Key takeaways

1
Executives’ top ambition for generative
AI adoption is driving increased
efficiency or productivity (72%), far
exceeding their interest in increasing

O
revenue (30%) or reducing costs (24%).
rganizations are starting the heavy lifting to

2
get real business value from generative AI. Strong data capabilities will be essential
As Arnab Chakraborty, chief responsible underpinnings to these aspirations, but
AI officer at Accenture, puts it, “2023 was only 22% of businesses consider their
the year when clients were amazed with data foundations “very ready” to support
generative AI and the possibilities. In 2024, we are generative AI applications today.
starting to see scaled implementations of responsible

3
generative AI programs.” The rise of AI exacerbates longstanding
challenges in data management—data
Some generative AI efforts remain modest. As Neil governance, security, and privacy (cited
Ward-Dutton, vice president for automation, analytics, by 59%), data quality and timeliness
and AI at IDC Europe, describes it, this is “a classic kind (53%), and data integration (48%)—and
of automation: making teams or individuals more may supply the urgency needed to finally
productive, getting rid of drudgery, and allowing people address them.
to deliver better results more quickly.”

Most companies, though, have much greater ambitions

for generative AI: they are looking to reshape how they
operate and what they sell. “AI is opening the door
Great expectations for generative AI
for organizations to gain
The expectation that generative AI could fundamentally insights from unstructured
upend business models and product offerings is driven
by the technology’s power to unlock vast amounts of
data that they simply
data that were previously inaccessible. “Eighty to couldn’t before.”
90% of the world’s data is unstructured,” says Baris
Gultekin, head of AI at AI data cloud company Snowflake. Baris Gultekin, Head of AI, Snowflake
“But what’s exciting is that AI is opening the door for
organizations to gain insights from this data that they In a poll conducted by MIT Technology Review Insights,
simply couldn’t before.” global executives were asked about the value they
hoped to derive from generative AI. Many say they are
prioritizing the technology’s ability to increase efficiency
and productivity (72%), increase market competitiveness

Methodology (55%), and drive better products and services (47%).

Few see the technology primarily as a driver of
increased revenue (30%) or reduced costs (24%),
MIT Technology Review Insights conducted a poll on
which is suggestive of executives’ loftier ambitions.
businesses’ data foundations for generative AI and
their ambitions and challenges when deploying and
Respondents’ top ambitions for generative AI seem to
scaling generative AI applications in May 2024. The
work hand in hand. More than half of companies say new
275+ executive respondents represent a broad range of
routes toward market competitiveness are one of their
industries and work at organizations across the globe.
top three goals, and the two likely paths they might take to
achieve this are increased efficiency and better products
or services.
 MIT Technology Review Insights 3

“There is no value without good business data.”

Neil Ward-Dutton, Vice President for Automation, Analytics, and AI, IDC Europe

For companies rolling out generative AI, these are not example. They, he says, are asking fundamental questions
necessarily distinct choices. Chakraborty sees a “thin about the technology’s power: “How can I use generative
line between efficiency and innovation” in current activity. AI to create new treatment pathways or to reimagine my
“We are starting to notice companies applying generative clinical trials process? Can I accelerate the drug discovery
AI agents for employees, and the use case is internal,” time frame from 10 years to five years to one?”
he says, but the time saved on mundane tasks allows
personnel to focus on customer service or more creative Data strategy underlies AI innovation
activities. Gultekin agrees. “We’re seeing innovation with Behind the diversity of ways in which respondents hope
customers building internal generative AI products that to secure value from generative AI looms one common-
unlock a lot of value,” he says. “They’re being built for ality: the need for enormous quantities of the business’s
productivity gains and efficiencies.” own data, accessibly stored and ready to use. Off-the-
shelf AI tools will not differentiate businesses when their
Chakraborty cites marketing campaigns as an example: adoption will soon be universal. For any enterprise AI
“The whole supply chain of creative input is getting use case, says Ward-Dutton, “there is no value without
re-imagined using the power of generative AI. That is good business data” of the company’s own.
obviously going to create new levels of efficiency, but
at the same time probably create innovation in the way
you bring new product ideas into the market.” Similarly,
Gultekin reports that a global technology conglomerate
and Snowflake customer has used AI to make “700,000
pages of research available to their team so that they Executive aspirations for generative
can ask questions and then increase the pace of their AI adoption
own innovation.” What are the primary types of value your organization
hope to achieve from its generative AI efforts?
First choice Second choice Third choice
The impact of generative AI on chatbots—in Gultekin’s
words, “the bread and butter of the recent AI cycle”— Increased efficiency or productivity
may be the best example. The rapid expansion in chatbot 36% 23% 13% 72%
capabilities using AI borders between the improvement
of an existing tool and creation of a new one. It is Increased market competitiveness
unsurprising, then, that 44% of respondents see improved 24% 14% 17% 55%

customer satisfaction as a way that generative AI will

Innovation in products or services
bring value.
12% 13% 22% 47%

A closer look at our survey results reflects this overlap Improved customer satisfaction
between productivity enhancement and product or 7% 18% 19% 44%
service innovation. Nearly one-third of respondents (30%)
included both increased productivity and innovation in Increased revenue
10% 9% 30%
the top three types of value they hope to achieve with 11%

generative AI. The first, in many cases, will serve as the

Faster decision-making
main route to the other.
6% 10% 13% 29%

But efficiency gains are not the only path to product Reduced costs
or service innovation. Some companies, Chakraborty 6% 10% 8% 24%

says, are “making big bets” on wholesale innovation with

Source: MIT Technology Review Insights poll, 2024
generative AI. He cites pharmaceutical companies as an
4  MIT Technology Review Insights

One of the great benefits of generative AI is that it can

make use of unstructured data, which many companies
Efficiency gains and innovation go
have in great quantities, such as wikis, documents, and
hand in hand
webpages. As Gultekin explains, “AI makes all of this
What are the primary types of value your organization
data much more useful because now you can unlock hopes to achieve from its generative AI efforts?
a massive amount of it that wasn’t easily accessible
before.” Moreover, humans can ask questions of their
data using natural language. This has a significant
impact on enterprises, helping them deliver insights
from previously untapped sources of information.
42% 30%
Selected just Selected
17%
Selected just
increased both innovation in
efficiency/ products or
This flood of newly accessible data, though, represents productivity services
both an opportunity and a competitive requirement.
Chakraborty describes the challenge: “To create Source: MIT Technology Review Insights poll, 2024

significant differentiation that becomes a competitive

asset, you need to start almost building your own large Data foundations are “somewhat ready”
language models, which means you need to train on for generative AI
massive amounts of data harnessed from within your To what extent are your organization’s data foundations
enterprise—from finance to HR to supply chain to ready to support generative AI applications?
marketing—and marry that with external data.”
Very
ready

22%
How ready, though, are most companies’ data estates
to support them in the race to generative AI value?
Somewhat
Gultekin puts it plainly: “The data foundation is at the core ready
of generative AI capabilities.” Data foundations cover a
broad collection of processes and assets involved in
the gathering, aggregation, storage, and accessibility
Very
unready 53%
of organizational data.
8% Somewhat
unready
Chakraborty identifies three specific data capabilities
necessary to support the effective deployment of
17%
Source: MIT Technology Review Insights poll, 2024
generative AI: the quality of the data; the ability to
integrate multiple sources of data; and the timely well, because data is very dynamic and the speed at which
democratization of data to relevant business users. it is getting created is magnifying every month.”
These are persistent difficulties, he notes, saying “all
have been issues that we have heard about for 10 years.” At first glance, most poll respondents seem positive about
These challenges will not easily be addressed this time, the state of their company data foundations. The majority
either, he warns: “We still will be 10 years from now, as rate their business “very ready” (22%) or “somewhat

“To create significant differentiation that becomes a

competitive asset, you need to start almost building your
own large language models, which means you need to
train on massive amounts of data harnessed from within
your enterprise and marry that with external data.”
Arnab Chakraborty, Chief Responsible AI Officer, Accenture
 MIT Technology Review Insights 5

ready” (53%) to support generative AI applications, while

about one-quarter acknowledge that their organization is
either “somewhat unready” (17%) or “very unready” (8%). The hallucination
This distribution is consistent with the experience of challenge
experts. Ward-Dutton, for example, says that, in his
Hallucinations—or the tendency of generative AI models
interactions with companies, “We consistently see a bell
to produce confident, but incorrect outputs—are a major
curve. The people at the front end have already made
challenge organizations face in deploying AI applications.
significant investments in things like data governance and
After all, no company wants to deploy a chatbot that makes
data quality. There are also a bunch of people still using,
false promises about its return policies to its customers or
basically, Excel for everything.”
an internal data tool that delivers unfounded insights to
its analysts.
Chakraborty warns that being “somewhat ready” could
conceal dangerous weaknesses. It might describe, for Baris Gultekin of Snowflake sees hallucinations as a
example, an organization that can integrate data sets major concern, and he has advocated for an insistence on
from across the business, but that also knows that higher accuracy rates in the next generation of AI-based
the underlying data is not fully clean. If so, he adds, tools. “For large enterprises, AI accuracy is imperative,
“Somewhat ready is not good enough. On accuracy and and providing users with trustworthy AI will be the leading
the quality of the data, 50% ready is not good enough. factor for businesses’ AI success at scale,” Gultekin
It needs to be higher to create trust.” He adds, “On the notes. Ensuring the quality and cleanliness of the data
ability to democratize that data to business users”— used to train AI models is of course an essential first step
to get users the data they need when they need it— to improving their outputs, he continues—“garbage in,
“somewhat ready is also not good enough.” garbage out” is a common refrain.

IDC’s research, Ward-Dutton says, shows that “somewhat Companies are also addressing hallucinations through
ready” is rarely “almost ready” in practice. The company’s improved techniques for ensuring accurate AI responses.
surveys have found that only 30 to 40% of businesses are Retrieval augmented generation (RAG), for example, is a
confident in their ability to perform each of the following technique used to ensure that AI outputs are grounded
data controls in their AI work: strictly control sensitive in verifiable outside data. Guardrails can impose firm
data with certainty that none will be leaked during model boundaries around the language or topics generative AI
training or use; manage use of third-party IP included in outputs can contain, and human and computer review of
the models they are using and ensure that their own IP outputs can iteratively improve an AI application’s results.
does not leak; and track and control how generative AI is
And sometimes it’s best for AI to simply acknowledge
interacting with their own internal data. His conclusion
that it doesn’t know. Gultekin says, “We’ve been focused
from such data is that “overall, readiness is pretty
on building products that know when to abstain from
equivocal.”
answering a question. LLMs are blissfully ignorant about
when they should not be answering, which is a dangerous
Nor is it all clear sailing for companies self-ranked “very
slope for businesses’ critical decision-making. We’ve
ready” by respondents. Instead, investment of resources
been investing in making systems that know when not
and time in their data foundations may just have made
to answer.”
clear their next set of challenges.

The main challenges of deploying generative AI at

scale are different for companies who say their data
foundations are “very ready.” Predictably, higher readiness
correlates with fewer challenges related to access to
scalable computing power, data silos and integration
issues, and data governance. In contrast, “very ready”
companies face more data quality issues, presumably
because better access to the information has revealed
6  MIT Technology Review Insights

its shortcomings. As Chakraborty says, “data-driven

transformation is not something where you can say, ‘I’ve A wide range of AI deployment
done it for the last two to three years, and now I’m done.’ and scaling challenges
Which of the following challenges does your organization
It’s a journey that will never be done.”
face when it comes to deploying AI at scale?

That said, generative AI’s benefits are becoming visible Data governance, security, or privacy
to those companies farther along the road. Gultekin 59%
reports that companies that have invested heavily in data
Data quality or timeliness
foundations are “all now reaping the benefits, because
53%
they are able to bring AI onto that data.” Ward-Dutton
adds that this deployment of generative AI represents a Costs or resource investment
business leap as well as a technological one. Companies
49%
that have invested in data governance and quality, he says,
“typically have also got to the point where they’re really Data silos or data integration challenges
engaging business people in how to manage data.” Now, 48%
the business people see, “‘Wow, this data actually has Access to scalable computing power
value for me.’”
25%

Roadblocks in the search for AI value Cloud readiness

Nothing good comes easy. Nearly all poll respondents 13%
acknowledge challenges in deploying generative AI.
None of these
Only 5% said that they did not face any of the challenges
listed in the question. 5%

Source: MIT Technology Review Insights poll, 2024

A majority of respondents (59%) cite data governance,
security, or privacy as a challenge. Nearly as many (53%)
say the same of data quality or timeliness. Meanwhile, It is unsurprising that data governance, security, or
nearly half face difficulties with cost (49%) and with privacy is the most widespread challenge. Corporate data
overcoming data integration challenges (48%). represent a company’s “crown jewels,” says Gultekin.
These are longstanding issues in data management, Deep concern about it is inevitable. Ward-Dutton adds
but the advent of AI complicates them further. It also, that, despite “quite good awareness at a high level” of the
as Ward-Dutton explains, increases the urgency of finally risks involved, “people are still flailing around a bit to find
grappling with them. “For a long time, many organizations the right mix of responses.”
have been able to limp along, or even do relatively well,
without being very good at data-related technical, cultural, Gultekin summarizes the many governance concerns
and organizational capabilities,” he says. “We’re seeing a surrounding generative AI: “At the foundational level, AI
lot having to catch up fast.” starts with securing and governing data. Companies want
 MIT Technology Review Insights 7

to bring AI closer to their data, ensuring that data security capability piece and the organizational and cultural one
and privacy are upheld. On top of that, customers want are two sides of the same coin.”
assurances that they are not liable for the data that the
large language model was trained with, and they want to Data silos and data integration remain a challenge for
ensure that their data does not get used to improve the nearly half of organizations. Gultekin notes that generative
model for others without their permission. These are all AI has heightened awareness of this problem as well.
table stakes and should be the industry standard.” By making large amounts of data all “much more useful
and much more accessible,” he says, “it is shedding a lot
Safe use of generative AI requires careful governance of light on the data infrastructure that is needed.” He
of its data sources. Chakraborty explains that the vast argues that investment in a single, standardized data
amount of public and company data used to train foundation across the organization will enable much
a large language model brings with it inherent risks. more powerful generative AI uses. It will also reduce
“You need a very surgical view in data governance,” he governance and security concerns: “When you keep
says. From sourcing data to creating outputs, companies your data in one place for one thing, another place for
require very strong data governance, including clarity another thing, governing and securing that data
on who owns the data and who stewards it. Adding to becomes really difficult,” he says.
the challenge, Gultekin explains, “is that the new data
generative AI makes accessible may not itself be Spending and resource decisions, including those needed
fully governed.” to shore up data foundations, are of course a challenge
when it comes to any technology investment. On the
Data quality and timeliness is another predictable concern. bright side, the cost of generative AI itself is decreasing
As Chakraborty puts it, “if your data quality is not clean, substantially. In particular, Gultekin reports, in recent
you are going to get all kinds of garbage.” Yet because months, enterprises have begun creating smaller LLMs
of the opaque way in which the technology works, this that remain extremely capable while being less expensive.
may not be obvious. This is where data lineage becomes “All these distilled models are still as good,” he says.
critical, ensuring transparency and traceability of data as “They are also very cost-effective and very good from
it moves through systems. For those scaling up generative a latency perspective.”
AI pilots, thereby giving the technology a greater role,
such quality issues magnify, adds Gultekin. As organizations feel increasing urgency to deploy
AI applications, they are realizing that their data is the
While IT departments are typically told to fix low-quality key to how quickly and effectively they can unlock new
data, the problem is caused where the information began value. Many will find that their generative AI ambitions
“in the first place: that typically is the business,” says are just castles in the air if they don’t have the right data
Ward-Dutton. Employees who understand the value of foundations to support them. A strong data strategy,
their data “are much more likely to care about the quality however, can guide them to surmount their data
of what they’re providing,” he says. “The technical or the challenges on the way to AI success.

“For a long time, many organizations have been able to

limp along, or even do relatively well, without being very
good at data-related technical, cultural, and organizational
capabilities. We’re seeing a lot having to catch up fast.”
Neil Ward-Dutton, Vice President for Automation, Analytics, and AI, IDC Europe
8  MIT Technology Review Insights

“Data strategies for AI leaders” is an executive briefing paper by MIT Technology Review Insights. We would like to thank
all participants as well as the sponsor, Snowflake. MIT Technology Review Insights has collected and reported on all
findings contained in this paper independently, regardless of participation or sponsorship. Teresa Elsey was the editor
of this report, and Nicola Crepaldi was the publisher.

About MIT Technology Review Insights

MIT Technology Review Insights is the custom publishing division of MIT Technology Review, the world’s
longest-running technology magazine, backed by the world’s foremost technology institution—producing
live events and research on the leading technology and business challenges of the day. Insights conducts
qualitative and quantitative research and analysis in the US and abroad and publishes a wide variety of
content, including articles, reports, infographics, videos, and podcasts. And through its growing MIT Technology
Review Global Insights Panel, Insights has unparalleled access to senior-level executives, innovators, and
entrepreneurs worldwide for surveys and in-depth interviews.

From the sponsor

Snowflake makes enterprise AI easy, efficient, and trusted. Thousands of companies around the globe,
including hundreds of the world’s largest, use Snowflake’s AI Data Cloud to share data, build applications, and
power their business with AI. The era of enterprise AI is here. Learn more at snowflake.com (NYSE: SNOW).

What is the AI Data Cloud?

The AI Data Cloud is a unified service used by over 10,000 companies to power their businesses with
data, AI, and applications. Snowflake’s AI Data Cloud consists of platform capabilities that support diverse data,
AI, and applications workloads, as well as content—the datasets, models, and apps themselves—that are
available to share and consume natively in the AI Data Cloud. The AI Data Cloud platform includes:
• Interoperable Storage: Snowflake-hosted and optimized storage (or extends to cloud data lakes)
• Elastic Compute: a unified engine supporting many languages and runtime needs
• Cortex AI: a managed AI layer including foundational LLMs, Chat API, and UI studio
• Cloud Services: infrastructure delivered as a service, including high availability and maintenance
• Snowgrid: a cross-region, cross-cloud network supporting data sharing and business continuity

Illustrations
Cover art and spot illustrations created with Adobe Stock.

While every effort has been taken to verify the accuracy of this information, MIT Technology Review Insights cannot accept any responsibility or liability for reliance by any person
on this report or any of the information, opinions, or conclusions set out in this report.

© Copyright MIT Technology Review Insights, 2024. All rights reserved.

MIT Technology Review Insights
www.technologyreview.com
[email protected]